Signal, Image and Video Processing Signal Processing

SPSD-YOLO: a novel paradigm for small-object detection in UAV imagery

Ru Xu, Zhang, Quan, Liu, Qingke
Springer Nature · 2026-11-13 · DOI 10.1007/s11760-026-05745-0

Abstract

Low-altitude aerial imagery contains objects that are small, densely distributed, partially occluded, and embedded in complex backgrounds, while UAV platforms impose strict computing limits. The proposed Sharp Perception Sensitive Edge Decoupling YOLO (SPSD-YOLO) follows a three-stage design: edge-preserving feature extraction, small-object feature reconstruction, and dynamic task decoupling. These stages respectively preserve fine boundaries, recover high-resolution representations, and adapt the detection head to scale, location, and task cues. On the VisDrone2019 validation set, the 2.3-M-parameter SPSD-YOLO $$_n$$ n obtains 44.3% mAP@50 and 26.9% mAP@50:95, improving over YOLO11n by 11.2 and 7.6 percentage points while using 11.5% fewer parameters. The results indicate a favorable accuracy-efficiency trade-off for resource-constrained edge inference.

Read at publisher →